Upload
truongkien
View
215
Download
0
Embed Size (px)
Citation preview
Testing hypotheses about compound stress assignment in English:
a corpus-based investigation
Ingo Plag, Gero Kunter, Sabine Lappe & Maria Braun
Universität Siegen
October 27, 2006
Corresponding author:
Ingo Plag English Linguistics Fachbereich 3 Universitaet Siegen Adolf-Reichwein-Str. 2 D-57068 Siegen http://www.uni-siegen.de/~engspra/ tel. +49-(0)271-740-2560 tel. +49-(0)271-740-2349 (secretary) fax tel. +49-(0)271-740-3246 e-mail: [email protected]
2
Testing hypotheses about compound stress assignment in English:
a corpus-based investigation
Abstract
This paper tests three factors that have been held to be responsible for the variable
stress behavior of noun-noun constructs in English: argument structure, semantics,
and analogy. In a large-scale investigation of some 4500 compounds extracted from
the CELEX lexical data base (Baayen et al. 1995), we show that traditional claims
about noun-noun stress cannot be upheld. Argument structure plays a role only with
synthetic compounds ending in the agentive suffix –er. The semantic categories and
relations assumed in the literature to trigger rightward stress do not show the
expected effects. As an alternative to the rule-based approaches, the data were
modeled computationally and probabilistically using a memory-based analogical
algorithm and logistic regression, respectively. It turns out that probabilistic models
and analogical algorithms are more successful in predicting stress assignment
correctly than any of the rules proposed in the literature. The behavior of the
analogical model suggests that the left and right constituent are the most important
factor in compound stress assignment. This is in line with recent findings on the
semi-regular behavior of compounds in other languages.
3
1. Introduction1
The present paper deals with an area of grammar which is more variable than
generally assumed: stress assignment in English noun-noun (NN) compounds. In
general, it has often been claimed that compounds tend to have a stress pattern that
is different from that of phrases. This is especially true for nominal compounds, the
class of compounds that is most productive (e.g. Plag 2003: 145). While phrases tend
to be stressed phrase-finally, compounds tend to be stressed on the first element. This
systematic difference is captured in the so-called nuclear stress and compound stress
rules (Chomsky and Halle 1968:17). Phonetic studies (e.g. Farnetani and Cosi 1988,
Ingram et al. 2003) have shown in addition that segmentally identical phrases and
compounds (such as bláckboard vs. black bóard) differ significantly not only in their
stress pattern, but also in length, with phrases being generally longer than the
corresponding compounds. While the compound stress rule apparently makes
correct predictions for what seems to be the majority of nominal compounds, it has
been pointed out, e.g. by Kingdon (1958), Fudge (1984), Liberman and Sproat (1992),
Bauer (1998), Olsen (2000, 2001), and Giegerich (2004), that there are also numerous
exceptions to the rule. Some of these forms are listed in (1). The most prominent
syllable is marked by an acute accent on the vowel.
(1) geologist-astrónomer apple píe scholar-áctivist
apricot crúmble Michigan hóspital Madison Ávenue
Boston márathon Penny Láne summer níght
aluminum fóil May flówers silk tíe
In view of this situation, the obvious question is how we can account for the
variability in stress assignment of noun-noun constructs. Basically, one finds three
kinds of hypotheses that are spelled out in the literature to different degrees of
explicitness. These hypotheses, which will be discussed in more detail shortly, refer
to either structural, semantic, or analogical factors that are held responsible for the
1 Acknowledgements to be added after review.
4
stress of NN-constructs. It has to be stated, however, that systematic empirical or
experimental work on the variability of compound stress is scarce, and the
hypotheses mentioned were develped the basis of data that had the following rather
questionable properties. First, the provenance of the data remained obscure. Authors
generally did not say where they took their data from. The selection of data does not
seem in any way systematic but more designed to prove the point of the respective
author. The second problem is that the amount of data is usually quite small, ranging
from only a handful of pertinent examples to a few hundred forms. The third
problem is that most of the studies do not discuss the details of their methodological
decisions, such as the assignment of particular examples to a given analytical
category.
In sum, there is a need for a large-scale empirical investigation of compound
stress variability using an independently gathered set of data. The present paper will
provide such a study. We will present the results of the investigation of all noun-
noun compounds (some 4500 types) extracted from the CELEX lexical data base
(Baayen et al. 1995). It will be shown that traditional claims about noun-noun stress
cannot be upheld. As an alternative to the rule-based approaches, we will model the
data computationally using a memory-based analogical algorithm, and
probabilistically using regression analysis. Overall, it turns out that the probabilistic
and analogical models are more successful in predicting stress assignment correctly
than any of the rules proposed in the literature. The behavior of the analogical model
suggests that the left and right constituent are the most important factor in
compound stress assignment.
Before we turn to the discussion of the hypotheses to be tested, a word is in
order with regard to the notorious problem of whether NN constructions should be
analyzed as compounds or phrases. Since we will use the CELEX lexical database,
this decision has already been taken by the compilers of that source. The constructs
that we investigate were considered words, hence compounds, by the compilers,
since CELEX only contains words, and not phrases.
The structure of the paper is as follows. Section 2 reviews the three hypotheses
on the variability of compound stress mentioned above. In section 3 we describe the
CELEX lexical data base and our data coding procedure, discussing the
5
methodological problems involved. Section 4 presents the results for the structural
hypothesis, section 5 for the semantic hypothesis and section 6 for the analogical
hypothesis. This is followed by the final discussion and conclusion in section 7.
2. Three hypotheses on stress assignment to compounds
Three types of approaches have been taken to account for the puzzling facts of
variable NN stress. The first is what Plag (2006) has called the ‘structural hypothesis’.
Proponents of this hypothesis (e.g. Bloomfield 1933, Lees 1963, Marchand 1969 or
Payne/Huddleston 2002) maintain that compounds are regularly left-stressed, and
that word combinations with rightward stress cannot be compounds, which raises
the question of what else such structures could be. One natural possibility is to
consider such forms to be phrases. However, such an approach would face the
problem of explaining why not all forms that have the same superficial structure, i.e.
NN, are phrases. Second, one would like to have independent criteria coinciding
with stress in order to say whether something is a lexical entity (i.e. a compound) or a
syntactic entity (i.e. a phrase). This is, however, often impossible: apart from stress
itself, there seems to be no independent argument for claiming that Mádison Street
should be a compound, whereas Madison Ávenue (or Madison Róad, for that matter)
should be a phrase. Both kinds of constructs seem to have the same internal
structure, both show the same meaning relationship between their respective
constituents, both are right-headed, and it is only in their stress patterns that they
differ. Spencer (2003) also argues that we find compounds with phrasal stress and
phrases with compound stress, and hence that stress is more related to lexicalization
patterns than to structural differences, a point taken up by Giegerich (2004, to be
discussed in more detail shortly). A final problem for the phrasal analysis is the fact
that the rightward stress pattern seems often triggered by analogy to other
combinations with the same rightward element. This can only happen if the forms on
which the analogy is based are stored in the mental lexicon. And storage in the
mental lexicon is something we would typically expect from words (i.e. compounds),
and only exceptionally from phrases (as in the case of jack-in-the-box).
6
Most recently, Giegerich (2004) has proposed a new variant of the structural
hypothesis. On the basis of the fact that in English syntax complements follow the
head, he argues that, due to the order of elements, complement-head structures like
trúck driver cannot be syntactic phrases, hence must be compounds, hence are left-
stressed. Modifier-head structures such as steel brídge display the same word order as
corresponding modifier-head phrases (cf. wooden brídge), hence are syntactic
structures and regularly right-stressed.2
This means, however, that many existing modifier-head structures are in fact
not stressed in the predicted way, since they are left-stressed (e.g. ópera glasses, táble
cloth). Such aberrant behavior, is, according to Giegerich, the result of lexicalization.
The problem with this idea is that lexicalization is, first, not a categorical notion, but
rather a gradual one, and, second, that it is not exactly clear how it can be decided
whether a given item is lexicalized or not. For compounds, four criteria come to
mind: frequency, spelling, semantic transparency, and phonological transparency. In
this study we will use frequency and spelling as indicators of lexicalization. Higher
frequency indicates a higher degree of lexicalization, and one-word spellings should
also be most prevalent with lexicalized compounds, while less lexicalized
compounds should prefer two-word spellings.
Lexicalization as an explanation for leftward stress makes interesting
predictions. With regard to corpora data, we should expect that the amount of
leftward-stressed compounds should vary according to frequency.3 Thus, we should
find more modifier-head structures with leftward stress among the more frequent
items. In addition, we would expect a higher proportion of left-stressed compounds
2 Giegerich characterizes modifier-head structures in terms of their lack of argument-predicate
semantics. We prefer the term ‘argument-head’ instead of ‘argument-predicate’ in the context of this
paper because of its parallelism with ‘modifier-head’. 3 Cf. Lipka’s definition, according to which lexicalization “is defined as the process by which complex
lexemes tend to become a single unit, with a specific content, through frequent use” (1994:2165, my
emphasis). Bauer (1983:51) mentions irregular stress assignment in English derivatives and Danish
compounds as prototypical cases of (phonological) lexicalization. See also Adams (1973:59), who
writes that “in established NPs which are used frequently and over a period of time the nucleus tends to
shift from the second to the first element although this does not always happen ” (our emphasis).
7
among those spelled as one word than among those spelled as two words, with
hyphenated compounds being somewhere in between.
Furthermore, the structural hypothesis predicts that we should never find
rightward stress among those NN constructs that exhibit complement-head order.
This is, however, not true, as pointed out by Giegerich himself, who cites Tory léader
as a counterexample. The structural hypothesis also entails that compounds with the
same rightward constituent exhibit different stress patterns, depending on whether
the leftward constituent is an argument or a modifier. Pairs such as yárd sale vs. bóok
sale (or trúck driver vs. Súnday driver) suggest that this prediction is probably not
always in accordance with the data. In such cases lexicalization would have to kick in
to explain leftward stress.
In any case, none of these predictions has ever been systematically tested
against larger amounts of data. In a recent experimental study, Plag (2006) found the
expected argument-structure effect, but no lexicalization effect (based on frequency).
Novel, i.e. newly invented, modifier-head compounds showed the same type of
variability in stress behavior as existing modifier-head compounds. Argument-head
compounds thus behaved as expected, while for modifier-head compounds the
hypothesis did not make the right predictions.
Before turning to the discussion of what we call the ‘semantic hypothesis’ we
would like to point out that what has been labeled ‘structural hypothesis’ is the
hypothesis that rests largely on the argument-modifier distinction. Although this
distinction clearly has strong semantic implications, there are, as pointed out above,
crucial structural facts that correlate with this distinction. This is our reason for
calling the hypothesis structural, athough the underlying distinction might be
semantic.
The second approach to variable compound stress is what can be called the
semantic hypothesis. A number of scholars have argued that words with rightward
stress such as those in (1) above are systematic exceptions to the compound stress
rule (e.g. Sampson 1980, Fudge 1984, Ladd 1984, Liberman and Sproat 1992, Olsen
2000, 2001, Spencer 2003). Although these authors differ slightly in details of their
respective approaches, they all argue that rightward prominence is restricted to only
a limited number of more or less well-defined types of meaning categories and
8
relationships. For example, compounds like geologist-astrónomer and scholar-áctivist
are copulative compounds, and these are uncontroversially and regularly right-
stressed.4 Other meaning relationships that are often, if not typically, accompanied
by rightward stress are temporal (e.g. summer níght), locative (e.g. Boston márathon),
and causative, the latter of which is usually paraphrased as ‘made of’ (as in aluminum
fóil, silk tíe), or ‘created by’ (as in a Shakespeare sónnet, a Mahler sýmphony). It is,
however, unclear how accurate the membership in a given class can really predict the
locus of stress. The leftward stress on súmmer school, súmmer camp or dáy job, for
example, violates Fudge’s (1984: 144ff.) generalization that NNs in which N1 refers to
a period or point of time are right-stressed. Furthermore, it is unclear how many, and
which, semantic classes should be set up to account for all the putative exceptions to
the compound stress rule (see also Bauer 1998:71 on this point). Finally, semantically
very similar compounds can behave differently in terms of stress assignment (Fífth
Street vs. Fifth Ávenue). And again, we have to state that, apart from the copulative
compounds (Olsen 2001) and compounds expressing an authorship relation (Plag
2006), detailed and systematic empirical studies are lacking for the classes postulated
to trigger rightward stress.
In the afore-mentioned experimental study by Plag (2006) it was tested
whether the semantic hypothesis makes the right predictions for compounds with an
authorship relation. Testing the authorship relation (as in Kauffmann sonata) against a
relation that is not predicted to trigger righthand stress (as in Twilight Sonata), it
turned out that the data show either no effect, or show an effect in the opposite
direction of what the semantic hypothesis would have predicted.
Note that we use the label ‘semantic hypothesis’ in this paper to refer to
approaches that set up semantic categories and correlate these with stress patterns.
Although these approaches actually never refer explicitly to the modifier-argument
distinction, the semantic categories that are alleged to produce rightward stress
would all involve modifier-head compounds, but never argument-head compounds.
4 Even this nice generalization has its (apparently very few) exceptions, for example mán-servant,
which is left-stressed.
9
Thus, structural and semantic hypothesis converge on the point that they expect
rightward stress to be largely restricted to modifier-head compounds.
Finally, a third approach can be taken which draws on the idea of analogy and
hypothesizes that stress assignment is generally based on analogy to existing NN
constructions in the mental lexicon. Plag (2003:139) mentions the textbook examples
of street vs. avenue compounds as a clear case of analogy. All street names involving
street as their right-hand constituent pattern alike in having leftward stress (e.g.
Óxford Street, Máin Street, Fóurth Street), while all combinations with, for example,
avenue as right-hand constituent pattern alike in having rightward stress (e.g. Fifth
Ávenue, Madison Ávenue). Schmerling (1971: 56) provides more examples of this kind,
arguing that many compounds choose their stress pattern in analogy to combinations
that have the same head, i.e. rightward constituent. It is, however, unclear how far
such an analogical approach can reach. Along similar lines, Spencer (2003: 331)
proposes that “stress patterns are in many cases determined by (admittedly vague)
semantic ‘constructions’ defined over collections of similar lexical entries.” In a
similar vein, Ladd (1984) proposes a destressing account of compound stress which
would explain the analogical effects triggered by the same rightward constituents as
basically semantico-pragmatic effects.
What is considered the effect of lexicalization in some approaches would
emerge naturally in an analogical system, in which existing (i.e. lexicalized)
compounds influence new (i.e. non-lexicalized) compounds to behave similarly. This
raises the question on which basis similarity could be computed (cf. also Liberman
and Sproat 1992: 176 on this point). In principle, any property could serve that
purpose, for example, the number of syllables of the right constituent, the semantic
properties of the left constituent, or, perhaps absurdly, the third segment of the left
constituent, or a combination of these. One rather simple assumption to start out
with is that it is the left or right constituent that is responsible for the choice of the
stress pattern. Given, for example, a set of compounds with the same right
constituent, we would first expect that the vast majority of items in that set are
stressed in a certain way, e.g. leftward, and that any novel form with that right-hand
constituent will also receive leftward stress. A more sophisticated analogical model
10
would incorporate of course more, and different types of linguistic information
(phonological, semantic, structural, frequential).
Plag (2006) found a robust effect of the right constituent on the stress of the
novel compounds used in the experiment, irrespective of the semantic relation
expressed by the compound. However, other potential factors playing a role in
analogy were not investigated in that study.
At this point a note is in order on the notion of analogy as used in different
traditions. The traditional notion of analogy has been rightly criticized by many
because it is difficult to see how any falsifiable prediction might be obtained with it.
For instance, in Goldberg and Jackendoff (2005: 475) we still find the statement that
‘analogy is notoriously difficult to constrain’. Recent work in computational
morphology has shown, however, that a formal, constrained, and computationally
tractable notion is available that offers new ways of understanding the ways in
which linguistic rules actually work. Such formal analogical models have been quite
successful in predicting both regular and irregular morphology in general, and
variable compound behavior in particular. For example, Krott and her collaborators
(Krott et al. 2001, Krott et al. 2002, Krott et al. 2004) analyzed the semi-regular
behavior of the linking morphemes in Dutch compounds in terms of analogy, using a
memory-based analogical learning algorithm (TiMBL, Daelemans et al. 2000).5 They
compared the algorithm’s performance with that of native speakers in an experiment
with novel compounds and found that the variable occurrence of the three linking
morphemes in Dutch compounds is much better accounted for by a dynamic
analogical mechanism than by traditional symbolic rules. Although the analogical
hypothesis has been evoked here and there (and quite informally) in treatments of
English compounds, it has never been tested empirically or formally modelled (cf.
Spencer’s above-cited remark on the vagueness of possible analogical sets).
To summarize, there are three reasonable hypotheses available to account for
the variability of NN stress, all of which are in some sense problematic and all of
which are still in need of serious empirical testing. One way to do this is to carry out
5 Analogical effects in compound interpretation have been shown to exist by Gagné and her
collaborators (e.g. Gagné and Shoben 1997, Gagné 2001).
11
experimental studies such as Plag’s (2006), in which the data can be carefully
controlled for the different potential factors involved. The present investigation takes
a different approach and uses the CELEX lexical database.
3. Methodology: general remarks
CELEX is a lexical database which contains data from German, (British) English and
Dutch, and has been successfully employed in linguistic and psycholinguistic
research, including compounds (e.g. Krott et al. 2001). Apart from orthographic
features, the CELEX database comprises representations of the phonological,
morphological, syntactic and frequency properties of lemmata. In this study we use
the English part of CELEX, which has been compiled on the basis of dictionary data
and text corpus data. The dictionary data come from the Oxford Advanced Learner's
Dictionary (1974, 41,000 lemmata) and from the Longman Dictionary of Contemporary
English (1978, 53,000 lemmata). The text corpus data come from the COBUILD
corpus, which contains 17.9 million word tokens. 92% of the word types attested in
COBUILD were incorporated into CELEX. The frequency information given in
CELEX is based on the COBUILD frequencies. Overall, CELEX contains lexical
information about 52,446 lemmata, which represent 160,594 word forms.
From the set of lemmata we selected all words that had two (and only two)
nouns as morphological constituents. This gave us a set of 4491 NN compounds. For
each of the compounds we extracted the information about stress and frequency. We
also coded each compound according to the categories held to be responsible for
stress assignment in the literature (and some more, to be discussed below). For those
variables where categorization proved to be problematic due to the ill-defined nature
of the categories mentioned in the literature, each compound was coded
independently by two raters and we analyzed only that subset of the data where the
two raters came up with the same categorization. Overall, three raters were
enganged in the coding, all of them holding both an MA and a PhD in English
linguistics.
12
To test the structural and semantic hypotheses we modeled the data
statistically using logistic regression, and compared the predictive accuracy of our
model with that of the two hypotheses. To test the analogical hypothesis we modeled
the data using a memory-based learner (TiMBL 5.1, Daelemans et al. 2004). Further
details of the methodologies employed will be discussed as we go along.
Before turning to the individual hypotheses let us take a first look at the data.
The overall distribution of stresses for the CELEX NN compounds is given in the
following bar graph.
Figure 1: Overall distribution of stress in NN compounds (N = 4491)
left right
stress position
num
ber o
f com
poun
ds
010
0020
0030
0040
00
We see that roughly 90 percent of the compounds in CELEX are given as left-
stressed, and 10 percent as right-stressed. In the following sections we will
investigate the nature of this variation in detail.
left stress right stressnumber of observations
4029 462
percentage 89.7 10.3
13
4. Testing the structural hypothesis
If we first take a look at the role of the argument structure distinction, we want to
take into account only those compounds where both ratings agreed. Thus for a
compound to be included here, both raters had to assign the same structural analysis,
for example ‘argument-head’. This reduces our data set to 4139 compounds. The
mosaic plot in figure 2 shows the distribution of stress by structure:
Figure 2: Distribution of stress by structure (N = 4139)
Mosaic plots represent the number of observations in each subset of the data as an
area. We can thus see that the majority of compounds has a modifier-head structure,
and that the proportion of right stresses is lower with argument-head compounds. A
chi-square analysis reveals that the difference between modifier-head and argument-
head compounds is statistically significant (χ2 = 8.55, df = 1, p = 0.003457, φ = 0.05).
Although the effect goes in the direction expected under the structural hypothesis,
the hypothesis that modifier-head structures be right-stressed is clearly falsified,
since of the 3806 modifier-head compounds, only 11 percent are right-stressed.
stress modifier- head
argument- head
left 3387 314
right 419 19
stre
ss p
ositi
on
modifier-head argument-head
left
right
14
In order to take a closer look at what is going on here we coded the
morphological makeup of the deverbal head nouns and investigated whether we
would find an interaction of suffix and argument structure. Under the structural
hypothesis we should expect that there be significant differences also between those
argument-head compounds and modifier-head compounds that share the same head
morpheme. The following table gives examples of the kinds of combinations we
found in the data:
Table 1:
morphology of head argument-head modifier-head
conversion fish slice side glance
-er squadron leader belly dancer
-ing trend setting sea-bathing
-ion blood transfusion Mercator projection
We also found a few heads that ended in the deverbal suffixes –age, –al, and –ance,
but these were too rare to be included in the statistical analysis. Overall, the heads of
683 items contained one of the suffixes shown in table 1 as the outermost suffix. The
distribution of stresses for this set of data is given in figure 3:
15
Figure 3: Interaction of head morphology, argument structure and stress
structure
mor
phol
ogy
argument-head modif ier-head
con
er
ing
ion
left right left right
An inspection of the plot already shows that for conversion, -ing, and –ion, the
distribution of stress does not seem to differ according to the modifier-argument
distinction. A logistic regression analysis of the interaction of argument structure and
righthand-head morpheme indeed reveals a significant effect only for those
compounds that have –er as their right-hand head morpheme (p = 0.0375, C = 0.723).
This restriction of the argument structure effect to –er compounds is the same as the
one recently found by Plag et al. (2006) in a study using an American speech corpus
(Boston University Radio Speech Corpus, Ostendorf et al. 1996). These findings can
be interpreted in such a way that the argument-structure effect hypothesized in the
literature is in fact an effect of only one particular subgroup of synthetic compounds,
those ending in -er. Not surprisingly, this is the subgroup that is almost exclusively
discussed in the literature, while the other subgroups are being largely ignored.
16
Let us compare the performance of our logistic regression model with that of
the stress assignment rule of the structural hypothesis. We used our logistic
regression model to calculate the probability of right stress for each compound on the
basis of the variables ‘argument structure’ and ‘morphology of the head’. If the
probablity of right stress was < 0.5 for a given item, we interpreted this item as left-
stressed, and as right-stressed if otherwise. These probabilistic predictions were then
compared to the stress positions found in CELEX – a match was counted as a correct
prediction. We also estimated the accuracy of the structural hypothesis by
categorically assigning left stress to argument-head compounds and right stress to
modifier-head compounds, and then comparing these stresses with the CELEX
stresses. The following table illustrates the methodology and gives the results for the
logistic regression model.
Table 2: accuracy of predictions, logistic regression model
Total correct predictions
incorrect predictions
prediction accuracy
CELEX has left stress 623 617 6 99.0%
CELEX has right stress 58 6 52 10.3%
Total 681 623 58 91.5%
The following table 3 compares the accuracies of the model and the structural
hypothesis:
Table 3: Accuracy of prediction, structural hypothesis
prediction regression model accuracy
structural hypothesis accuracy
of left stresses 99.0% 46.9%
of right stresses 10.3% 72.4%
total 91.5% 49.0%
The overall accuracy is far better for the logistic regression model (91.5% vs. 49.0%).
The structural hypothesis captures exsisting right stresses to a relatively high degree
17
of 72.4% (as against only 10.3% correct right stresses in regression), but it also assigns
incorrectly right stresses to the large number of modifier-head compounds that are
actually left-stressed. In view of this problem the obvious escape hatch for the
structural hypothesis is lexicalization, to which we now turn.
We will first investigate lexicalization using frequency. The problem with the
CELEX frequencies is that many compounds in CELEX are taken from the
dictionaries and are not attested in the COBUILD corpus, so that they are listed with
a frequency of zero. We took only those compounds whose frequency is larger than
zero, which gives us still 2118 observations. For this subset we do not find the
expected lexicalization effect. Compounds with leftward stress do not have a
significantly higher CELEX frequency (t (2128) < 1) than compounds with rightward
stress. This is illustrated in the boxplot in figure 4:
Figure 4: Stress position and CELEX frequency (all items with frequency > 0)
There is no interaction between stress position and argument structure (F (1, 2126) <
1), which means that neither the modifier-head compounds nor the argument head
compounds show the expected effect. This means that, contra the structural
hypothesis, there is no lexicalization effect observable (if we use the CELEX
frequencies).
18
In order to overcome the difficulty of having lost half of the data points due to
lacking CELEX frequencies, we also took a look at the log Google frequencies of all
compounds. First we checked the reliability of the Google frequencies6 by correlating
them with the CELEX frequencies, which come, as mentioned above, from a
controlled corpus. The reliability of the Google frequencies was confirmed by their
strong correlation (Spearman’s ρ = 0.511, p < 0.05) with the CELEX frequencies. Only
for items with a log CELEX frequency lower than log (11) = 2.398 does the relation
deviate from linearity, as shown by the scatterplot smoother in figure 5:7
Figure 5: Google log frequency by CELEX frequency
6 See Plag (2006: 159) for a detailed discussion of the problems involved in using Google for
investigating compound frequencies. 7 Given the wide usage of newsletter in an Internet context, and the resulting extremely high Google
frequency, this item was excluded from the comparison.
19
Having established the reliability of the Google frequencies, we repeated the
frequency analysis from above with the whole set of compounds in CELEX, now
with their Google frequencies. Using the Google frequencies does not improve the
situation for the structural hypothesis, however. We still do not find the expected
effect (t (4469) < 1). Consider figure 6 for illustration:
Figure 6: Stress position by Google frequency
Again, there is no interaction between stress position and argument structure (F (1,
4466) = 2.801, p > 0.05), which means that neither the modifier-head compounds nor
the argument head compounds show the expected effect. Contra the structural
hypothesis, there is no lexicalization effect observable.
Finally, we investigate the idea that spelling may be an indicator of
lexicalization, and can thus be used to test the structural hypothesis. It can be
assumed that one-word spellings should be most prevalent with lexicalized
compounds, while less lexicalized compounds should prefer two-word spellings.
According to the structural hypothesis we would therefore expect the proportion of
right stresses to be highest among the two-word compounds, lower among the
hyphenated compounds, and lowest among the compounds spelled as one
orthographic word. Indeed this effect can be found (χ2 = 512.08, df = 2, p < 0.01).
left right
510
1520
log
Goo
gle
frequ
ency
20
Figure 7 nicely illustrates the trend: the tighter the orthography, the more likely
becomes leftward stress.
Figure 7: Stress by spelling
spelling
stre
ss p
ositi
on
one w ord hyphenated tw o w ords
left
right
However, when examining the difference between argument-head and modifier-
head compounds, an analysis of deviance of a generalized linear model revealed no
significant reduction of the residuals for the interaction between spelling and stress
position (logit, df = 2, p = 0.1), so that, contra the hypothesis, we find a general
lexicalization effect, but not one that is restricted to modifier-head compounds.
To summarize our results for the structural hypothesis, we can say that this
hypothesis is not very successful in predicting compound stress. The argument
structure effect is restricted to compounds ending in –er, and the predicted
lexicalization effect is not measurable (when using frequency as a correlate) and is
not restricted to modifier-head compounds (when using spelling). Let us turn to the
semantic hypothesis and see whether it fares better.
21
5. Testing the semantic hypothesis
As mentioned in section 2 we often find claims concerning rightward stress
assignment which are based on semantic considerations. In general these
considerations refer either to the semantic relationship between the two compound
constituents, or to the properties of individual compound constituents or of the
compound as a whole. For ease of reference we will refer to the former set of
semantic entities as ‘(semantic) relations’, and to the latter set of semantic entities as
‘(semantic) categories’.
The literature predicts rightward stress explicitly for the following semantic
categories (e.g. Fudge 1984: 144ff, Liberman & Sproat 1992, Zwicky 1986); ‘N1’ refers
to the left constituent, ‘N2’ to the right constituent):
(2) N1 refers to a period or point in time (as in night bird)
N2 is a geographical term (lee shore),
N2 is a type of thoroughfare (chain bridge)
The compound is a proper noun (Union Jack)
N1 is a proper noun (Achilles tendon)
In addition, the literature claims that rigthward stress is triggered by the following
semantic relations (e.g. Fudge 1984: 144ff, Liberman & Sproat 1992):
(3) N2 DURING N1 (harvest festival)
N2 IS LOCATED AT N1 (promenade concert)
N2 IS MADE OF N1 (tin hat)
N1 MAKES N2 (worm hole)
There are a number of methodological and theoretical problems with testing these
claims. First of all, the semantic categories and semantic relations mentioned in the
literature (such as ‚N1 is a material‘, ‘N2 is located at N1’) seem generally ill-defined.
22
Second, items are often ambiguous, i.e. they show more than one relation.8 Third, on
a theoretical level it is unclear how many and what kinds of relations and categories
would be expected to play a role. There may be many more (or less) than the eight
categories and relations mentioned above that have an effect on stress assignment.
In order to deal with, if not solve, these problems we used a set of 18 semantic
relations that are more or less established as useful in studies of compound
interpretation. The bulk of these relations come from Levi (1978), a seminal work on
compound semantics, whose relations have since been employed in many linguistic
(e.g. Liberman & Sproat 1992) and, more recently, psycholinguistic studies of
compound structure, stress and meaning (cf., for example, Gagné & Shoben 1997,
Gagné 2001). Levi’s catalogue contains fewer than our 18 relations, but we felt that
some additions were necessary, especially to ensure the possibility of reciprocal
relations. For example, Levi’s list has a relation N2 USES N1, but no relation N1
USES N2. In such cases we added the missing relation to our set of relations to be
coded. Furthermore, we added a few categories that we felt were missing from her
set, such as N2 IS NAMED AFTER N1. In (4) we present the final list of our relations.
The relations are expressed by supposedly language-independent predicates that
link the concepts denoted by the two constituents (see Levi 1978 for discussion).
8 Cf., for example, worm hole, which could also be interpreted in a locative sense.
23
(4) List of semantic relations coded, illustrated with one example each
Semantic relation example
1. N2 CAUSES N1 teargas
2. N1 CAUSES N2 heat rash
3. N2 HAS N1 stock market
4. N1 HAS N2 lung power
5. N2 MAKES N1 silkworm
6. N1 MAKES N2 firelight
7. N2 IS MADE OF N1 potato crisp
8. N2 USES N1 water mill
9. N1 USES N2 handbrake
10. N1 IS N2 child prodigy
11. N1 IS LIKE N2 kettle drum
12. N2 FOR N1 travel agency
13. N2 ABOUT N1 mortality table
14. N2 IS LOCATED
AT/IN/...
N1 garden party
15. N1 IS LOCATED
AT/IN/...
N2 taxi stand
16. N2 DURING N1 night watch
17. N2 IS NAMED
AFTER
N1 Wellington boot
18. OTHER schoolfellow
Some of the categories proved especially difficult to code consistently, so that
additional guidelines were developed. These concerned mainly the interpretation of
the predicates CAUSE, MAKE, and IS. CAUSE was pertinent in cases where a cause
(denoted by the one constituent) triggers an effect (denoted by the other constituent),
while MAKE was coded in cases of purposeful creation or of production. IS
subsumes three cases, the first being that the left constituent denotes a subset of the
denotation of the right constituent (poison gas), the second being that left and right
24
constituents are not in a subset-superset relation and IS works in both directions (girl-
friend), the third being same-level appositional compounds (owner-driver).
Given that compounds in English are in principle ambiguous, a compound could
be assigned multiple relationships. As mentioned above, each compound was coded
by two independent raters and only those compounds were analyzed that were
assigned to the same category by the two raters.
The data were subjected to two separate logistic regression analyses, the first
using the five semantic categories referring to compound constituents or the
compound as a whole, the second using the 18 relations as predictors. Let us first
look at the categories referring to compound constituents or the compound as a
whole. The five panels in figure 8 illustrate the distributions of stresses for these
types of compound:
Figure 8: Stress position by semantic category
The logistic regression analysis with the five categories as predictors shows that for
three of the five categories (upper panels: ‘N1 refers to a point of time’ and ‘N2 is a
N1 refers to POINT OF TIME?
no yes
left
right
N2 is a GEOGRAPHICAL TERM?
no yes
left
right
N2 is a THOROUGHFARE?
no yes
left
right
N1 is a PROPER NOUN?
no yes
left
right
Compound is a PROPER NOUN?
no yes
left
right
25
geographical term’, and ‘N2 is a thoroughfare’) we do not find the predicted effect.
For the two other categories (‘compound is a proper noun’, and ‘N1 is a proper
noun’) we do find the expected effect, i.e. compounds with these categories have a
higher proportion of right-stresses than other compounds (‘compound is a proper
noun’: p = 0.004, and ‘N1 is a proper noun’: p < 0.001). However, the majority of the
pertinent compounds are still left-stressed, so that the overall predictive power of the
model is extremely low (C = 0.551).
Let us turn to the regression analysis of the effect of semantic relations on
stress assignment. Of the four relations predicted to favor rightward stress only three
could be tested (N2 DURING N1, N2 LOCATED AT N1, N2 IS MADE OF N1),
because there were only two pertinent cases for the relation N1 MAKES N2. In the
analysis of deviance of a logistic regression model, two relations did not have the
predicted effect (N2 LOCATED AT N1: p = 0.53, N2 DURING N1: p = 0.97), while
the third, N2 IS MADE OF N1, did show a significant effect in the right direction (p <
0.001). In addition, our model shows significant rightward stress effects also for N1
IS N2 (p < 0.001) and N2 IS NAMED AFTER N1 (p < 0.001), and a significant
leftward stress effect for N2 FOR N1 (p < 0.001). The overall power of the model is
not too impressive (C = 0.772). Figure 9 shows the distributions:
Figure 9: Stress position by semantic relation
N2 DURING N1
no yes
left
right
N2 LO CATED AT/IN N1
no yes
left
right
N2 IS M ADE O F N1
no yes
left
right
N1 IS N2
no yes
left
right
N2 FO R N1
no yes
left
right
N2 ÍS NAM ED AFTER N1
no yes
left
right
26
The findings for semantic relation thus closely resemble those for semantic categories
discussed above, in that only a subset of the proposed variables show an effect in the
expected direction of rightward stress, but that the majority of the pertinent
compounds are still left-stressed.
If we combine all semantic relations and all semantic categories in one
combined logistic regression model, the four relations N1 IS MADE OF N2, N1 IS N2,
N2 FOR N1, N2 IS NAMED AFTER N1 remain significant, while of the semantic
categories only ‘The compound is a proper noun’ remains in the model.9 All
significant predictors increase the likelihood of rightward stress, only N2 FOR N1
increases left-stress. A comparison of the performance of the final regression model
and the semantic hypothesis (based on the categorical implementation of the
categories and relations found in the literature) again reveals a better overall
accuracy (the two models differ significantly: χ2 = 179.984, df = 3, p < 0.01) for the
regression model, with the categorical rules (again) overpredicting right stresses and
underpredicting left stresses. The regression model, on the other hand, has more
trouble accounting for the occurrence of right stress in the data set. Table 4 gives the
pertinent figures:
Table 4: Accuracy of prediction, semantic hypothesis
prediction regression model accuracy
semantic hypothesis accuracy
of left stresses 98.3% 85.0%
of right stresses 15.2% 30.0%
total 89.0% 78.7%
We may now turn to the final hypothesis, analogy.
9 The significances in the analysis of deviance for N1 IS MADE OF N2, N1 IS N2, N2 FOR N1, and N2
IS NAMED AFTER N1 are all p < 0.001, for ‘The compound is a proper noun’ we get p = 0.011.
27
6. Testing the analogical hypothesis
In general terms, the analogical hypothesis claims that stress in compounds is
determined by the stress pattern of the majority of similar instances that are stored in
memory. For example, a given word carpet beater would be assigned leftward stress
because the putatively most similar exemplar stored in memory (say, e.g., eggbeater),
has leftward stress. The crucial problem is of course how to measure the similarity
between compounds. In the present study, we used TiMBL 5.1 (Daelemans et al.
2004) as a computational algorithm to test the analogical hypothesis.
TiMBL computes similarities by counting identical values of the
variables that encode the properties of the items in the lexicon and the input. This
works as follows. For every compound in the lexicon, we have coded its semantic
and structural properties and its stress. When a new compound comes in and needs
to be assigned stress, the new compound is compared in all its properties with all
exemplars in the lexicon. The algorithm selects a set of so-called nearest neighbors
which contains only those compounds that are most similar to the input. The
algorithm then assigns the kind of stress that is most frequent among the nearest
neighbors. Figure 10 illustrates the procedure using only five predictor categories
(left constituent, right constituent, argument structure, morphology, and semantics).
In the example, oil painting is assigned leftward stress since both compounds in the
set of nearest neighbours are left-stressed.
28
Figure 10: Stress assignment by TiMBL
All claims about analogy that can be found in the literature make reference to the left
and right constituents of compounds (see again section 2 for details and examples).
In order to be able to specifically test these claims (in addition to testing the potential
analogical effects of argument-structural, semantic and morphological factors) we
used only those compounds whose left or right constituent occurred more than once
in the corpus. In other words, if we wanted to test the effect of consituent families,
we had to have only those compounds for which the pertinent information was
available. While this procedure reduced our data set to 2643 items, it made it possible
for us to test whether the absence vs. the presence of the constituent family
information would make a difference in the accuracy of predictions.
Let us see how good the algorithm is at taking the right decisions for the 2643
compounds. Table 5 gives the correct and incorrect predictions for each type of
stress. It also compares the predictive accuracy reached by TiMBL with that of a
regression model based on this subset of the data.
evaluation of input against nearest neighbours
INPUT OUTPUT
action, painting, noarg, -ing, semcat1, stress: leftfinger, painting, noarg, -ing, semcat1, stress: leftwall, painting, arg, -ing, semcat1, stress: leftcountry, party, noarg, nosuff, semcat2, stress: leftcottage, hospital, noarg, nosuff, semcat3,life, work, noarg, semcat2, nosuff, stress: right ...
} stress left: 2xstress right: 0x
oil, painting, noarg, -ing, semcat1 stress left
INSTANCE-BASED
MEMORY
SET OF NEAREST NEIGHBOURS
stress: right
action, painting, noarg, -ing, semcat1, stress: leftfinger, painting, noarg, -ing, semcat1, stress: left
29
Table 5: Accuracy of prediction, analogical hypothesis
prediction regression model accuracy
TiMBL’s Accuracy
of left stresses 100% 99.0%
of right stresses 0.6% 21.2%
total 94.1% 94.4%
We can see that TiMBL, like our regression models, overpredicts leftward stress and
underpredicts rightward stress. If we compare TiMBL’s performance with that of the
rules and models discussed in the previous two section, we have to state that TiMBL
has the best overall performance of all, but considerable problems with predicting
rightward stresses. The logistic regression analysis, which, like TiMBL, makes use of
all predictor variables, does not differ significantly in its overall accuracy from
TiMBL. It detects 100% of the left stresses, but notably only 0.6% of the right stresses,
which leads to an overall accuracy of 94.1%.
The important question is of course which kinds of features prove to be most
important for the analogical algorithm. Given the claims in the literature, what is of
particular interest is the contribution of the constituent family information. The
above figure of 94.4% accurracy has been achieved by taking all kinds of feature into
account. If we take each set of features (structural, morphological, semantic, and
constituent family) separately, i.e. if we ignore all other features while testing one set
of features, we, quite surprisingly, always arrive at around 94% accurarcy. Thus, any
given set of features is as good a predictor as any other set. If we do the opposite and
use all sets of features but one, basically the same picture emerges. The performance
does not drop, with the significant exception of the factor ‘constituent family’. If we
leave out the information on the left or right constituent the accuracy rate drops
significantly (left constituent: χ2 = 7.425, p= 0.0064, right constituent: χ2 = 4.834, p =
0.0279, left and right constituent: χ2 = 4.834, p= 0.0279).
This result shows that left and right constituent indeed add valuable
information to the classfication task in an analogical model. Our findings thus
constitute robust evidence for the influence of constituent families on compound
stress assignment. Importantly, this evidence is not based on small sets of hand-
30
picked forms that supposedly show the validity of that hypothesis, but emerges
through the formal analysis of a large number of independently gathered data
points.
7. Discussion and conclusion
In this paper we have presented the first large-scale investigation of compound stress
in English based on independently available data. Overall the study has shown that
existing hypotheses about compound stress are not able to explain the variability in
the data. This is in line with recent experimental and speech corpus-based studies
(such as Plag 2006 and Plag et al. 2006).
In particular, the idea that argument structure plays a role in compound stress
assignment has been shown to only hold for compounds ending in the agentive
suffix –er, and not for compounds featuring other right-hand head morphemes (such
as –ion, -ing, or conversion). With regard to lexicalization effects, we found a clear
interaction of spelling and stress, with one-word compounds exhibiting almost
exclusively leftward stress. However, we did not find a lexicalization effect based on
frequency data, which runs counter to the spelling results. How can this discrepancy
be explained?
An answer to this question can be found if we look at the compounds spelled
as two words. First, we should note that the two-word compounds are the smallest
subset of compounds in CELEX (cf. again figure 7 above). Second, apart from two
exceptions, all of the 1268 two-word compounds in CELEX have a frequency of zero,
i.e. they all have been taken from the two dictionaries, and not from the COBUILD
corpus. This means that the (presumably very many) two-word compounds that
occur in the texts of the COBUILD corpus were simply not sampled for the CELEX
data base. For practical reasons of data base creation, only orthographic words, i.e.
continuous letter strings between two spaces, were sampled from the COBUILD
31
corpus.10 Since we can expect that the proportion of non-lexicalized compounds is
highest among two-word compounds, we have to state that CELEX has in general a
bias towards lexicalized compounds. This would also explain the rather high amount
of 90% left stresses in the data.11
The fact that CELEX has this bias towards lexicalized compounds may seem
disappointing, but we have to recall that – apart from Giegerich (2004) – none of the
hypotheses found in the literature is explicit about the role of lexicalization. In fact,
many of the examples cited in the pertinent literature to back up the respective
claims, are indeed well-known, hence lexicalized, compounds. Therefore, even a
database that has a lexicalization bias such as CELEX is an appropriate testing
ground for these theories.
With regard to the semantic hypothesis we have shown that only few
predictions are borne out by the facts, that many claims do not hold, and that it is
possible to find new effects. Finally, the analogical modeling of the data showed that
the constituent families play an important role in stress assignment. This supports
pertinent claims in the literature, which so far have rested on a few pertinent
examples only.
An overall comparison of models shows that analogical modeling is most
successful with the data. Table 6 below combines tables 3 through 5 for convenience.
We refer to the accuracy reached by the application of the categorical rules of the
structural and semantic hypotheses as ‘hypothesis-based accuracy’ :
10 That this is indeed the case has been confirmed by Harald Baayen (p.c.), one of the authors of
CELEX. The two exceptional items station wagon and India rubber are probably not from the COBUILD
corpus, but received their frequency of 3 due to an error. 11 It is presently unclear in what proportions left and right stresses occur in English compounds. In a
recent perception experiment using a random sample of compounds from the Boston University
Radio Speech Corpus, Kunter & Plag (2006) found, roughly, two thirds left stresses and one third right
stresses (type-wise). Plag et al. (2006) even find a majority of right-stressed compounds among their
4400 compounds, counting token-wise. It remains to be shown what kind of a distribution can be
regarded as representative of the language as a whole.
32
Table 6: Accuracy of predictions across hypotheses and models
structural factors semantic factors all factors prediction regression
model accuracy
hypothesis-based accuracy
regression model accuracy
hypothesis-based accuracy
regression model accuracy
TiMBL’s accuracy
of left stresses
99.0% 46.9% 98.3% 85.0% 100,0% 99.0%
of right stresses
10.3% 72.4% 15.2% 30.0% 0.6% 21.2%
total 91.5% 49.0% 89.0% 78.7% 94.1 94.4%
What these comparisons show us is that rule-based models cannot adequately cope
with the variability of the data. Probabilistic and analogical models have higher
overall accuracy rates, even though they are not good at detecting rightward stress.
Rule-based approaches account better for the -- in the CELEX data -- much less
frequent rightward stresses, while at the same time overgeneralizing rightward stress
incorrectly to many compounds with leftward stress. The overall results are in line
with the findings of recent studies of compounds in other languages (e.g. Krott et al.
2001, 2002, 2004), which have also shown that variable compound behavior is best
accounted for by probabilistic or analogical models, instead of rule-based ones.
Future studies of compound stress will have to show whether this general
assumption about the organization of compounds in the mental lexicon still holds if
more non-lexicalized data are factored in.
Finally, the present study has found additional evidence for the importance of
the constituent family in explaining compound behavior. Krott et al. (2001, 2002,
2004) have already shown that the constituent family has significant influence on the
choice of the linking morpheme in Dutch compounds, and Gagné (2001) provided
evidence that the constituent family has an effect on compound interpretation. Our
study now demonstrates that such effects also pertain to the phonology of
compounds.
33
References Adams, V. (1973). An introduction to Modern English word-formation. London: Longman. Baayen, H. R. H., R. Piepenbrock, and L. Gulikers (1995). The CELEX lexical database (CD-ROM).
Philadelphia: Linguistic Data Consortium, University of Pennsylvania. Bauer, L. (1983). English word-formation. Cambridge: CUP. Bauer, L. (1998). When is a sequence of two nouns a compound in English? English Language and
Linguistics 2: 65–86. Bloomfield, L. (1933). Language. Chicago: Holt. Chomsky, N. & M. Halle (1968). The sound pattern of English. New York: Harper and Row. Daelemans, W., J. Zavrel, K. van der Sloot, and A. van den Bosch (2004). TiMBL: Tilburg Memory Based
Learner, version 5.1, Reference Guide. ILK Technical Report 04-02, available from http://ilk.uvt.nl/downloads/pub/papers/ilk0402.pdf.
Farnetani, E. & P. Cosi (1988). English compound versus non-compound noun phrases in discourse: An acoustic and perceptual study. Language and Speech 31: 157-180.
Fudge, E. C. (1984). English word-stress. London: George Allen & Unwin. Gagné, Ch. & E. Shoben (1997). Influence of thematic relations on the comprehension of modifier-
noun combinations. Journal of Experimental Psychology: Learning, Memory and Cognition 23: 71-87.
Gagné, Ch. (2001). Relation and lexical priming during the interpretation of noun-noun combinations. Journal of Experimental Psychology: Learning, Memory and Cognition 27: 236-254.
Giegerich, H. (2004). Compound or phrase? English noun-plus-noun constructions and the stress criterion. English Language and Linguistics 8: 1–24.
Goldberg, A. & R. Jackendoff (2005). The end result(ative). Language 81.2: 474-477. Ingram, J. , T. A. T. Nguyen & R. Pensalfini (2003). An acoustic analysis of compound and phrasal
stress patterns in Australian English. Submitted for publication. Kingdon, R. (1958). The groundwork of English stress. London, etc.: Longmans, Green & Co. Krott, A. , H. Baayen & R. Schreuder (2001). Analogy in morphology: Modeling the choice of linking
morphemes in Dutch. Linguistics 39: 51-93. Krott, A. , P. Hagoort & H. Baayen (2004). Sublexical units and supralexical combinatorics: The case of
Dutch interfixes in visual processing. Language and Cognitive Processes 19: 453-471. Krott, A. , R. Schreuder & H. Baayen (2002). Analogical hierarchy: Exemplar-based modeling of linkers
in Dutch noun-noun compounds. In Skousen, R. , D. Lonsdale & D. S. Parkinson (eds.) Analogical modeling: An exemplar-based approach to language. Amsterdam: Benjamins. 181-206.
Kunter, G. & I. Plag (2006). What is compound stress? Paper presented at the University of Edinburgh, May 2006.
Ladd, D. R. (1984). English compound stress. In Gibbon, Dafydd & Helmut Richter (eds.) Intonation, accent and rhythm. Berlin: Mouton de Gruyter, 253-266.
Lees, R. B. (1963). The grammar of English nominalizations. The Hague: Mouton. Levi, J. N. (1978) The syntax and semantics of complex nominals. New York: Academic Press. Liberman, M. & R. Sproat (1992). The stress and structure of modified noun phrases in English. In Sag,
I. A. & A. Szabolcsi (eds.) Lexical Matters. Stanford: Center for the Study of Language and Information. 131-181.
Lipka, L. (1994). Lexicalization and institutionalization. In Adger, R. E. (ed.) The encyclopedia of language and linguistics. Oxford: Pergamon Press. 2164-2167.
Marchand, H. (1969). The categories and types of present-day English word-formation. 2nd ed. München: Beck.
Olsen, S. (2000). Compounding and stress in English: A closer look at the boundary between morphology and syntax. Linguistische Berichte 181: 55-69.
Olsen, S. (2001). Copulative compounds: A closer look at the interface between syntax and morphology. In Booij, G. E. & J. van Marle (eds.) Yearbook of Morphology 2000. Dordrecht/Boston/London: Kluwer. 279-320.
Ostendorf, M., P. Price, & S. Shattuck-Hufnagel (1996). Boston University Radio Speech Corpus. Philadelphia: Linguistic Data Consortium, University of Pennsylvania.Payne, J. , and R. Huddleston (2002). Nouns and noun phrases. In Huddleston, R. & G. K. Pullum, The Cambridge grammar of the English language. Cambridge: Cambridge University Press. 323–524.
34
Plag, I. (2003). Word-formation in English. Cambridge Textbooks in Linguistics. Cambridge: Cambridge University Press.
Plag, I. (2006). The variability of compound stress in English: structural, semantic and analogical factors, English Language and Linguistics 10.1, 143-172.
Plag, I., G. Kunter, S. Lappe & M. Braun (2006) Variable stress assignment in noun-noun compounds: new evidence from corpora, Directions in English Language Studies, University of Manchester, 6-8 April 2006.
Sampson, R. (1980). Stress in English N+N phrases: A further complicating factor. English Studies 61: 264-270.
Schmerling, S. F. (1971). A stress mess. Studies in the Linguistic Sciences 1: 52-65. Spencer, A. (2003). Does English have productive compounding? In Booij. G. E., J. DeCesaris, A. Ralli
& S. Scalise (eds.). Topics in morphology. Selected papers from the third Mediterranean morphology meeting (Barcelona, September 20—22, 2001). Barcelona: Institut Universitari de Lingüística Applicada, Universtitat Pompeu Fabra. 329—341.
Zwicky, A. M. (1986) Forestress and afterstress. Ohio State University Working Papers in Linguistics 32, 46-62.